{
  "id": 226708,
  "title": "18th Place Solution",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/226708",
  "author_name": "Miyatti",
  "post_date": "2021-03-17T11:42:29.145000",
  "votes": 29,
  "comment_count": 21,
  "views": 0,
  "content": "<p>Hi all! Thank you for hosting and organizing such a great competition. I learned many things from this competition.</p>\n<h1>Summary</h1>\n<ul>\n<li>As a starting point, I used <a href=\"https://www.kaggle.com/ammarali32\" target=\"_blank\">@ammarali32</a> and <a href=\"https://www.kaggle.com/underwearfitting\" target=\"_blank\">@underwearfitting</a> pretrained model. Thank you for sharing!!</li>\n<li>3rd Training Stages Strategy:<ul>\n<li>1st: basemodel + classifier for all class with light augmentations</li>\n<li>2nd: basemodel + classifier for each class with light augmentations</li>\n<li>3rd: basemodel + classifier for each class with heavy augmentations</li></ul></li>\n<li>As a heavy augmentation, I tried to use \"annotation-based mixcut\"<ul>\n<li>To increase data variation, randomly erase the annotated area and its label, and fill with non-label image</li></ul></li>\n<li>To avoid the Notebook Runtime Error, I measure runtime many times, then I think I could achieve 8hours 59mins runtime by giving up TTA for fold10 of Model3.<ul>\n<li>I was so lucky to avoid errors!</li></ul></li>\n</ul>\n<p>I believe I could have tried many other ideas like going deeply inside external dataset, but I competed solo and did not have enough time and GPU resources.</p>\n<p>Finally I got public 0.970(22nd) and private 0.973(18th). </p>\n<p><img src=\"https://pbs.twimg.com/media/EwrSCJZUYAEQCW3?format=jpg&amp;name=large\" alt=\"\"></p>",
  "messages": [
    {
      "id": 1242076,
      "postDate": "2021-03-17T11:42:29.147Z",
      "content": "<p>Hi all! Thank you for hosting and organizing such a great competition. I learned many things from this competition.</p>\n<h1>Summary</h1>\n<ul>\n<li>As a starting point, I used <a href=\"https://www.kaggle.com/ammarali32\" target=\"_blank\">@ammarali32</a> and <a href=\"https://www.kaggle.com/underwearfitting\" target=\"_blank\">@underwearfitting</a> pretrained model. Thank you for sharing!!</li>\n<li>3rd Training Stages Strategy:<ul>\n<li>1st: basemodel + classifier for all class with light augmentations</li>\n<li>2nd: basemodel + classifier for each class with light augmentations</li>\n<li>3rd: basemodel + classifier for each class with heavy augmentations</li></ul></li>\n<li>As a heavy augmentation, I tried to use \"annotation-based mixcut\"<ul>\n<li>To increase data variation, randomly erase the annotated area and its label, and fill with non-label image</li></ul></li>\n<li>To avoid the Notebook Runtime Error, I measure runtime many times, then I think I could achieve 8hours 59mins runtime by giving up TTA for fold10 of Model3.<ul>\n<li>I was so lucky to avoid errors!</li></ul></li>\n</ul>\n<p>I believe I could have tried many other ideas like going deeply inside external dataset, but I competed solo and did not have enough time and GPU resources.</p>\n<p>Finally I got public 0.970(22nd) and private 0.973(18th). </p>\n<p><img src=\"https://pbs.twimg.com/media/EwrSCJZUYAEQCW3?format=jpg&amp;name=large\" alt=\"\"></p>",
      "rawMarkdown": "Hi all! Thank you for hosting and organizing such a great competition. I learned many things from this competition.\n\n# Summary\n- As a starting point, I used @ammarali32 and @underwearfitting pretrained model. Thank you for sharing!!\n- 3rd Training Stages Strategy:\n    - 1st: basemodel + classifier for all class with light augmentations\n    - 2nd: basemodel + classifier for each class with light augmentations\n    - 3rd: basemodel + classifier for each class with heavy augmentations\n- As a heavy augmentation, I tried to use \"annotation-based mixcut\"\n    - To increase data variation, randomly erase the annotated area and its label, and fill with non-label image\n- To avoid the Notebook Runtime Error, I measure runtime many times, then I think I could achieve 8hours 59mins runtime by giving up TTA for fold10 of Model3.\n     - I was so lucky to avoid errors!\n\nI believe I could have tried many other ideas like going deeply inside external dataset, but I competed solo and did not have enough time and GPU resources.\n\nFinally I got public 0.970(22nd) and private 0.973(18th). \n\n![](https://pbs.twimg.com/media/EwrSCJZUYAEQCW3?format=jpg&name=large)",
      "votes": 29
    },
    {
      "id": 1242696,
      "postDate": "2021-03-17T18:39:22.850Z",
      "content": "<p>Impressive solo finish, especially with such a high ressource demanding competition. Congratz!</p>",
      "rawMarkdown": "Impressive solo finish, especially with such a high ressource demanding competition. Congratz!",
      "votes": 1,
      "replies": [
        {
          "id": 1243021,
          "postDate": "2021-03-18T00:52:27.347Z",
          "content": "<p>Exactly! We need more GPUs…</p>",
          "rawMarkdown": "Exactly! We need more GPUs...",
          "votes": 1
        }
      ]
    },
    {
      "id": 1242346,
      "postDate": "2021-03-17T14:58:51.837Z",
      "content": "<p>Congrats !! great work and results. and big thanks for your support.</p>",
      "rawMarkdown": "Congrats !! great work and results. and big thanks for your support.",
      "votes": 1,
      "replies": [
        {
          "id": 1242390,
          "postDate": "2021-03-17T15:27:06.333Z",
          "content": "<p>I really appreciate your effort and contribution. You are my hero.</p>",
          "rawMarkdown": "I really appreciate your effort and contribution. You are my hero.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1242278,
      "postDate": "2021-03-17T14:09:41.770Z",
      "content": "<p>Congratulations, Wow 8h 59m inference. lots of model ensembles. Thanks for sharing your idea. </p>",
      "rawMarkdown": "Congratulations, Wow 8h 59m inference. lots of model ensembles. Thanks for sharing your idea. ",
      "votes": 1,
      "replies": [
        {
          "id": 1242321,
          "postDate": "2021-03-17T14:43:33.807Z",
          "content": "<p>Thank you! I have had trouble in avoiding run time error…</p>",
          "rawMarkdown": "Thank you! I have had trouble in avoiding run time error..."
        }
      ]
    },
    {
      "id": 1242087,
      "postDate": "2021-03-17T11:49:44.517Z",
      "content": "<p><a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> Thanks for sharing great detailed solution writeup . Congratulations on Silver Finish </p>",
      "rawMarkdown": "@yoshitaka1105 Thanks for sharing great detailed solution writeup . Congratulations on Silver Finish ",
      "votes": 1,
      "replies": [
        {
          "id": 1242325,
          "postDate": "2021-03-17T14:45:07.940Z",
          "content": "<p>Thank you for much!</p>",
          "rawMarkdown": "Thank you for much!"
        }
      ]
    },
    {
      "id": 1242239,
      "postDate": "2021-03-17T13:46:32.203Z",
      "content": "<p>Congrats on solo strongly silver finish <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> </p>",
      "rawMarkdown": "Congrats on solo strongly silver finish @yoshitaka1105 ",
      "votes": 2,
      "replies": [
        {
          "id": 1242308,
          "postDate": "2021-03-17T14:37:24.527Z",
          "content": "<p>Thank you for your comment!</p>",
          "rawMarkdown": "Thank you for your comment!"
        }
      ]
    },
    {
      "id": 1242192,
      "postDate": "2021-03-17T13:16:07.180Z",
      "content": "<p>8hour59 minutes. This is the highlight hahaha </p>",
      "rawMarkdown": "8hour59 minutes. This is the highlight hahaha ",
      "votes": 2,
      "replies": [
        {
          "id": 1242324,
          "postDate": "2021-03-17T14:44:48.073Z",
          "content": "<p>Exactly! I added 20sec job to my submission notebook, then run time error happened…</p>",
          "rawMarkdown": "Exactly! I added 20sec job to my submission notebook, then run time error happened...",
          "votes": 1
        }
      ]
    },
    {
      "id": 1242103,
      "postDate": "2021-03-17T11:58:40.447Z",
      "content": "<p>Congratulation!  <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a>  Thank you for sharing.</p>\n<p>I want to check my understanding :<br>\nso only 1st: basemodel is a multi-label classification, but 2nd: basemodel &amp; 3rd: basemodel are not or you made 11 models (one for each class)?</p>\n<p>I am a bit confused about &gt; classifier for each class</p>\n<blockquote>\n  <p>8hours 59mins runtime !    </p>\n</blockquote>\n<p>wow :)</p>",
      "rawMarkdown": "Congratulation!  @yoshitaka1105  Thank you for sharing.\n\nI want to check my understanding :\nso only 1st: basemodel is a multi-label classification, but 2nd: basemodel & 3rd: basemodel are not or you made 11 models (one for each class)?\n\nI am a bit confused about > classifier for each class\n\n> 8hours 59mins runtime !    \n\n wow :)\n",
      "votes": 2,
      "replies": [
        {
          "id": 1242317,
          "postDate": "2021-03-17T14:42:05.013Z",
          "content": "<p>Sorry for your confusion. 2nd and 3rd model have \"multi-head\" classifier to each class. Technically it is one model, but can be considered as 11 models. </p>",
          "rawMarkdown": "Sorry for your confusion. 2nd and 3rd model have \"multi-head\" classifier to each class. Technically it is one model, but can be considered as 11 models. ",
          "votes": 1
        },
        {
          "id": 1243846,
          "postDate": "2021-03-18T14:26:15.283Z",
          "content": "<p>Thank you very much.</p>",
          "rawMarkdown": "Thank you very much."
        }
      ]
    },
    {
      "id": 1242081,
      "postDate": "2021-03-17T11:45:42.600Z",
      "content": "<p>Oh my god! my mind blows with nothing<br>\nCongratulations! for the well deserved solution</p>",
      "rawMarkdown": "Oh my god! my mind blows with nothing\nCongratulations! for the well deserved solution",
      "votes": 2,
      "replies": [
        {
          "id": 1242320,
          "postDate": "2021-03-17T14:42:56.193Z",
          "content": "<p>Thank you for kind comment and competing with me. It was really exciting.</p>",
          "rawMarkdown": "Thank you for kind comment and competing with me. It was really exciting.",
          "votes": 1
        }
      ]
    },
    {
      "id": 1243155,
      "postDate": "2021-03-18T03:55:31.303Z",
      "content": "<p>Congratulations <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> - \"annotation-based mixcut\" is a novel idea here. Were your models all 512 image size? <br>\nOvercoming Notebook Runtime Error and GPU limits was a real achievement here!  </p>",
      "rawMarkdown": "Congratulations @yoshitaka1105 - \"annotation-based mixcut\" is a novel idea here. Were your models all 512 image size? \nOvercoming Notebook Runtime Error and GPU limits was a real achievement here!  ",
      "replies": [
        {
          "id": 1243614,
          "postDate": "2021-03-18T11:14:10.623Z",
          "content": "<p>Thank you! My all models are for 640x640 image size.</p>",
          "rawMarkdown": "Thank you! My all models are for 640x640 image size."
        }
      ]
    },
    {
      "id": 1246823,
      "postDate": "2021-03-21T06:21:25.267Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true,
      "replies": [
        {
          "id": 1247919,
          "postDate": "2021-03-22T07:27:13.783Z",
          "content": "<p>Thank you!</p>",
          "rawMarkdown": "Thank you!"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 1242696,
      "author_name": "Theo Viel",
      "author_url": "",
      "post_date": "2021-03-17T18:39:22.850000",
      "content": "<p>Impressive solo finish, especially with such a high ressource demanding competition. Congratz!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1243021,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-18T00:52:27.347000",
          "content": "<p>Exactly! We need more GPUs…</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1242346,
      "author_name": "ammarali32",
      "author_url": "",
      "post_date": "2021-03-17T14:58:51.837000",
      "content": "<p>Congrats !! great work and results. and big thanks for your support.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1242390,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T15:27:06.333000",
          "content": "<p>I really appreciate your effort and contribution. You are my hero.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1242278,
      "author_name": "Md. Masud Rana",
      "author_url": "",
      "post_date": "2021-03-17T14:09:41.770000",
      "content": "<p>Congratulations, Wow 8h 59m inference. lots of model ensembles. Thanks for sharing your idea. </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1242321,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:43:33.807000",
          "content": "<p>Thank you! I have had trouble in avoiding run time error…</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1242087,
      "author_name": "Tensor Girl",
      "author_url": "",
      "post_date": "2021-03-17T11:49:44.517000",
      "content": "<p><a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> Thanks for sharing great detailed solution writeup . Congratulations on Silver Finish </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1242325,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:45:07.940000",
          "content": "<p>Thank you for much!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1242239,
      "author_name": "KhanhVD",
      "author_url": "",
      "post_date": "2021-03-17T13:46:32.203000",
      "content": "<p>Congrats on solo strongly silver finish <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> </p>",
      "votes": 2,
      "replies": [
        {
          "id": 1242308,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:37:24.527000",
          "content": "<p>Thank you for your comment!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1242192,
      "author_name": "gao-hongnan",
      "author_url": "",
      "post_date": "2021-03-17T13:16:07.180000",
      "content": "<p>8hour59 minutes. This is the highlight hahaha </p>",
      "votes": 2,
      "replies": [
        {
          "id": 1242324,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:44:48.073000",
          "content": "<p>Exactly! I added 20sec job to my submission notebook, then run time error happened…</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1242103,
      "author_name": "Faisal Alsrheed",
      "author_url": "",
      "post_date": "2021-03-17T11:58:40.447000",
      "content": "<p>Congratulation!  <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a>  Thank you for sharing.</p>\n<p>I want to check my understanding :<br>\nso only 1st: basemodel is a multi-label classification, but 2nd: basemodel &amp; 3rd: basemodel are not or you made 11 models (one for each class)?</p>\n<p>I am a bit confused about &gt; classifier for each class</p>\n<blockquote>\n  <p>8hours 59mins runtime !    </p>\n</blockquote>\n<p>wow :)</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1242317,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:42:05.013000",
          "content": "<p>Sorry for your confusion. 2nd and 3rd model have \"multi-head\" classifier to each class. Technically it is one model, but can be considered as 11 models. </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1243846,
          "author_name": "Faisal Alsrheed",
          "author_url": "",
          "post_date": "2021-03-18T14:26:15.283000",
          "content": "<p>Thank you very much.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1242081,
      "author_name": "Mohammed Rizin V K",
      "author_url": "",
      "post_date": "2021-03-17T11:45:42.600000",
      "content": "<p>Oh my god! my mind blows with nothing<br>\nCongratulations! for the well deserved solution</p>",
      "votes": 2,
      "replies": [
        {
          "id": 1242320,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-17T14:42:56.193000",
          "content": "<p>Thank you for kind comment and competing with me. It was really exciting.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 1243155,
      "author_name": "something4kag",
      "author_url": "",
      "post_date": "2021-03-18T03:55:31.303000",
      "content": "<p>Congratulations <a href=\"https://www.kaggle.com/yoshitaka1105\" target=\"_blank\">@yoshitaka1105</a> - \"annotation-based mixcut\" is a novel idea here. Were your models all 512 image size? <br>\nOvercoming Notebook Runtime Error and GPU limits was a real achievement here!  </p>",
      "votes": 0,
      "replies": [
        {
          "id": 1243614,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-18T11:14:10.623000",
          "content": "<p>Thank you! My all models are for 640x640 image size.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1246823,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-03-21T06:21:25.267000",
      "content": "",
      "votes": 1,
      "replies": [
        {
          "id": 1247919,
          "author_name": "Miyatti",
          "author_url": "",
          "post_date": "2021-03-22T07:27:13.783000",
          "content": "<p>Thank you!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1242076": "Hi all! Thank you for hosting and organizing such a great competition. I learned many things from this competition.\n\n# Summary\n- As a starting point, I used @ammarali32 and @underwearfitting pretrained model. Thank you for sharing!!\n- 3rd Training Stages Strategy:\n    - 1st: basemodel + classifier for all class with light augmentations\n    - 2nd: basemodel + classifier for each class with light augmentations\n    - 3rd: basemodel + classifier for each class with heavy augmentations\n- As a heavy augmentation, I tried to use \"annotation-based mixcut\"\n    - To increase data variation, randomly erase the annotated area and its label, and fill with non-label image\n- To avoid the Notebook Runtime Error, I measure runtime many times, then I think I could achieve 8hours 59mins runtime by giving up TTA for fold10 of Model3.\n     - I was so lucky to avoid errors!\n\nI believe I could have tried many other ideas like going deeply inside external dataset, but I competed solo and did not have enough time and GPU resources.\n\nFinally I got public 0.970(22nd) and private 0.973(18th). \n\n![](https://pbs.twimg.com/media/EwrSCJZUYAEQCW3?format=jpg&name=large)",
    "1242696": "Impressive solo finish, especially with such a high ressource demanding competition. Congratz!",
    "1242346": "Congrats !! great work and results. and big thanks for your support.",
    "1242278": "Congratulations, Wow 8h 59m inference. lots of model ensembles. Thanks for sharing your idea. ",
    "1242087": "@yoshitaka1105 Thanks for sharing great detailed solution writeup . Congratulations on Silver Finish ",
    "1242239": "Congrats on solo strongly silver finish @yoshitaka1105 ",
    "1242192": "8hour59 minutes. This is the highlight hahaha ",
    "1242103": "Congratulation!  @yoshitaka1105  Thank you for sharing.\n\nI want to check my understanding :\nso only 1st: basemodel is a multi-label classification, but 2nd: basemodel & 3rd: basemodel are not or you made 11 models (one for each class)?\n\nI am a bit confused about > classifier for each class\n\n> 8hours 59mins runtime !    \n\n wow :)\n",
    "1242081": "Oh my god! my mind blows with nothing\nCongratulations! for the well deserved solution",
    "1243155": "Congratulations @yoshitaka1105 - \"annotation-based mixcut\" is a novel idea here. Were your models all 512 image size? \nOvercoming Notebook Runtime Error and GPU limits was a real achievement here!  ",
    "1246823": ""
  }
}